Close the accuracy-gate gaps that let a broken model be promoted - #1216

Merged

ZhengGong-amd merged 14 commits into

mainfrom

feature/leixin/eval-context-feasibility

Aug 20, 2026

Commits on Aug 17, 2026

Stop charging an unrunnable accuracy gate to the kernel

Zeng
and
cursoragent
committed
800a43f
Use the repository's MIT SPDX header on the new test file

Zeng
and
cursoragent
committed
a6d0703

Commits on Aug 18, 2026

Merge remote-tracking branch 'origin/main' into feature/leixin/eval-context-feasibility
Zeng
committed
b4d399d
Gate the serving-config lanes that could promote a broken model

Zeng
and
cursoragent
committed
1ae7431
Read the champion's verdict off the stack instead of the explore ledger

Zeng
and
cursoragent
committed
82e351a

Commits on Aug 19, 2026

Read the replay's accuracy from the round that measures it

Zeng
and
cursoragent
committed
e14f5fb
Record why a replay could not be judged, not just that it wasn't

Zeng
and
cursoragent
committed
7da5138
Merge origin/main into feature/leixin/eval-context-feasibility

Zeng
and
cursoragent
committed
7de70a6
Apply absolute accuracy floor when warm replay has no baseline reference.

Zeng
and
cursoragent
committed
e47a064

Commits on Aug 20, 2026

ci: retrigger Read the Docs build

Zeng
and
cursoragent
committed
d49a5ac
Merge remote-tracking branch 'origin/main' into feature/leixin/eval-context-feasibility
Zeng
committed
23881c5
Merge remote-tracking branch 'origin/main' into feature/leixin/eval-context-feasibility
Zeng
committed
c46541e
Do not report a parser that never read a file as an eval that ran.

Zeng
and
cursoragent
committed
6eeb9ff
Let --no-eval outrank the replay's forced warmup eval.

Zeng
and
cursoragent
committed
ee44084